An MCP server to have agents use Overleaf for free.
Once it's connected, an agent can:
- create or duplicate projects
- list, read, write, and edit
.texfiles - compile a project and pull the PDF back locally
If you dont have uv installed, install it.
It needs your Overleaf session cookie to do anything, so grab that first:
- Sign in at overleaf.com.
- Open DevTools -> Network, reload the project list, click any
overleaf.comrequest, and copy the fullCookierequest header.
Copy over this for the OVERLEAF_SESSION environment variable in the MCP server.
claude mcp add overleaf -e OVERLEAF_SESSION="overleaf_session2=s%3A...; GCLB=..." -- uvx overleaf-connectOpen Settings -> Developer -> Edit Config and add an overleaf entry under mcpServers (keep any servers you already have):
{
"mcpServers": {
"overleaf": {
"command": "uvx",
"args": ["overleaf-connect"],
"env": {
"OVERLEAF_SESSION": "overleaf_session2=s%3A...; GCLB=..."
}
}
}
}Afterwards, restart Claude Desktop to reload the MCP servers and get Overleaf connected.
Overleaf issues a fresh cookie on every request. After the first success the server saves the rotating value to:
- Windows:
%LOCALAPPDATA%\overleaf-connect\session.json - macOS / Linux:
~/.overleaf-connect/session.json
The OVERLEAF_SESSION environment variable is treated as a seed - the server will automatically refresh the session whenever requests are done to Overleaf, and store it. You only need to change it in the future if the cookie has expired from not using it.
OVERLEAF_HOME_DIR=... # directory that holds session.json
OVERLEAF_SESSION_FILE=... # exact session.json path
OVERLEAF_CACHE_DIR=... # compiled-PDF cache
OVERLEAF_OUTPUT_DIR=... # allowed download_pdf destination root
# (default: ~/Downloads/overleaf-connect, or
# ~/Documents/overleaf-connect, else ~/overleaf-connect)
| Tool | What it does |
|---|---|
status |
check if session cookie works |
list_projects |
id, name, last updated |
create_project |
blank project |
duplicate_project |
clone an existing project |
list_files |
paths in a project |
read_file |
text contents of a file (full=true for files over 100k chars) |
write_file |
whole-file create/replace |
edit_file |
exact old/new string replace in a file |
compile_project |
Overleaf compile -> cached output.pdf + parsed errors |
download_pdf |
compile and copy the PDF under OVERLEAF_OUTPUT_DIR |
trash_project |
move a project to Overleaf trash |
Writes replace the whole file, so write_file refuses to write back a file that
was only read in part - read it again with full=true, or use edit_file.
Don't edit a file while someone is typing in it in the browser.